Descriptors, not photographs
A well-designed system does not store a gallery of employee photos and compare pictures. It runs the enrolment image through a neural network that outputs a fixed-length vector – typically a few hundred numbers – describing the geometry of the face. Only this descriptor is kept, usually encrypted. At punch time the live frame is converted the same way and the distance between the two vectors is measured; below a threshold, it is the same person.
This matters for privacy and for accuracy. The descriptor cannot be reversed into a recognisable photo, which simplifies compliance with the DPDP Act 2023. And because the comparison is mathematical rather than pixel-based, moderate changes in beard, spectacles or lighting still match. Read how face recognition attendance works for the full pipeline.
Terminal vs phone
A wall-mounted face terminal suits a single gate with heavy footfall: infrared or 3D cameras, consistent lighting, sub-second matching and no dependence on employee phones. It costs hardware per site and cannot follow a guard to a new client location.
Phone-based face attendance uses the employee's own device or a shared tablet in kiosk mode. It adds a GPS stamp and geofence check for free and works at any site the moment the roster changes. The trade-offs are camera quality on budget Android phones and the need for stronger liveness detection, since an attacker controls the device.
- Choose a terminal for one gate, 200+ people, fixed lighting
- Choose phone-based face for multi-site, mobile or contract workforces
- Insist on liveness detection in both cases; a printed photo should fail
- Confirm descriptors, not images, are stored, and where (India region preferred)
What affects accuracy in Indian conditions
Lighting is the main factor: a guard at a gate at 20:00 with a street light behind him produces a silhouette. Good apps guide the user to face the light, and enrolment should be done in similar conditions to daily use. Masks, helmets with visors, and heavy religious head coverings that shadow the face reduce match confidence; the fix is a slightly relaxed threshold with a second factor (geofence or supervisor approval) rather than forcing removal.
Enrolment quality decides everything downstream. Take 3–5 frames, front-facing, no sunglasses, and re-enrol if someone's appearance changes materially. Set the threshold conservatively – false rejects cost a retry; false accepts cost you the anti-proxy benefit you bought the system for.
A facility-management firm deploys 48 housekeeping staff across six housing societies in Noida. Each marks in on the app at 07:00 by facing the camera; the match runs in about a second and a liveness prompt asks for a blink. The punch stores descriptor confidence, GPS point and geofence result. On a month with 1,152 expected punches, supervisors review only the 9 that failed geofence.
Attend Mitra's face recognition attendance stores encrypted face descriptors rather than photos, runs a liveness check on every punch, and combines the result with GPS, geofencing and mock-location flagging on Android. It works on the employee's own phone, on a shared kiosk tablet, or alongside biometric terminals at fixed gates.